Update README.md

README.md

@@ -21,6 +21,25 @@ Try it in your browser without any setup:
 
 Upload your own PCB image or pick from built-in test images. The confidence threshold is adjustable in the sidebar.
 
+
+## Dataset: DsPCBSD+
+
+DsPCBSD+ is a 2024 open dataset for PCB copper-layer defect detection, captured by a professional AOI system (AGLE'OL AOI-100 V8, 16K camera, controlled LED lighting). Nine defect categories, annotated at instance level:
+
+| Code | Defect | Description |
+|---|---|---|
+| SH | Short | Conductive bridge between two traces |
+| SP | Spur | Anomalous copper spike from a trace |
+| SC | Spurious Copper | Unwanted copper on board surface |
+| OP | Open Circuit | Break in a conductive trace |
+| MB | Mouse Bite | Edge notch or chip in the trace |
+| HB | Hole Breakout | Damage to material around a drill hole |
+| CS | Conductor Scratch | Scratch mark on a conductive trace |
+| CFO | Conductor Foreign Object | Foreign particle on a trace |
+| BMFO | Base Material Foreign Object | Foreign particle in substrate material |
+
+> S. Lv et al., "A dataset for deep learning based detection of printed circuit board surface defect," *Scientific Data*, vol. 11, no. 1, p. 811, 2024. https://doi.org/10.1038/s41597-024-03656-8
+
 ## Model
 
 ### Architecture: RT-DETRv4 X
@@ -79,8 +98,6 @@ Ground truth vs. predictions at conf≥0.3. The model correctly localises all fo
 | AP large (>96² px) | 0.754 |
 | AR @ maxDets=100 | 0.686 |
 
-![COCO detection metrics — AP/AR across IoU thresholds and object sizes, plus per-category AP](images/4_coco_metrics.png)
-
 ### Per-Class AP @ IoU=0.50:0.95
 
 | Class | Full Name | AP |
@@ -96,6 +113,8 @@ Ground truth vs. predictions at conf≥0.3. The model correctly localises all fo
 | BMFO | Base Material Foreign Object | 0.478 |
 | Mean | | 0.522 |
 
+![COCO detection metrics — AP/AR across IoU thresholds and object sizes, plus per-category AP](images/4_coco_metrics.png)
+
 ### Per-Class F1 @ Conf≥0.5, IoU≥0.5
 
 | Class | Precision | Recall | F1 |
@@ -130,31 +149,6 @@ Raw counts (left) and row-normalised rates (right). HB is the strongest class (0
 
 RT-DETRv4 X matches or exceeds all compared methods at only 10 epochs on a single T4. PCB-FS, for comparison, used 100+ epochs with more resources.
 
-## Dataset: DsPCBSD+
-
-DsPCBSD+ is a 2024 open dataset for PCB copper-layer defect detection, captured by a professional AOI system (AGLE'OL AOI-100 V8, 16K camera, controlled LED lighting). Nine defect categories, annotated at instance level:
-
-| Code | Defect | Description |
-|---|---|---|
-| SH | Short | Conductive bridge between two traces |
-| SP | Spur | Anomalous copper spike from a trace |
-| SC | Spurious Copper | Unwanted copper on board surface |
-| OP | Open Circuit | Break in a conductive trace |
-| MB | Mouse Bite | Edge notch or chip in the trace |
-| HB | Hole Breakout | Damage to material around a drill hole |
-| CS | Conductor Scratch | Scratch mark on a conductive trace |
-| CFO | Conductor Foreign Object | Foreign particle on a trace |
-| BMFO | Base Material Foreign Object | Foreign particle in substrate material |
-
-> S. Lv et al., "A dataset for deep learning based detection of printed circuit board surface defect," *Scientific Data*, vol. 11, no. 1, p. 811, 2024. https://doi.org/10.1038/s41597-024-03656-8
-
-## Files
-
-| File | Description |
-|---|---|
-| `last.pth` | Fine-tuned checkpoint (EMA weights) |
-| `test_image/` | 12 example PCB images for inference testing |
-
 ## Running Locally
 
 Clone the Space and install dependencies: